cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

laobrasuca
Journeyman III

Suggest Feature you want in AMD APP

 

This has been fixed already.  CLInfo info will not crash in OpenCL 1.0 platform. Instead it will exit with error message.

 

yes, you are right, I used the wrong word, my bad. It will not crash, it will lead to an error message indeed. What I was trying to say is that this error can be avoided if you don't use the CL_VERSION_1_1 macro as the ultimate reference for the opencl version for all platforms. A 'if/else' on the CL_PLATFORM_VERSION could avoid this instead of the "#ifdef CL_VERSION_1_1".

0 Likes
rubens
Journeyman III

Suggest Feature you want in AMD APP

What about say something about simultaneous writes to same memory position.

like CUDA especification does:"It states that when when warp makes multiple writes to the same memory locations, at least one of the writes will sucessed. Which one that is in undefined."

 

0 Likes
Meteorhead
Challenger

Suggest Feature you want in AMD APP

It has never occured to me that I would've wanted to rely on something like that to work. And I would disencourage anyone from using methods like this. Puting statements such as:

if(get_local_id(0) == 0) {...}

is surely slower than letting all of the threads write the same variable and know that one of them will succeed. However writing code that "accidently" works on GPU is not portable code and most likely does not work on other vendor's cards, not to mention CPUs.

OpenCL is there to create a general layer of doing paralell computation. Going around specification is alright for short term hacking, but should (and can be!) avoided at all times.

0 Likes
mrbpix
Journeyman III

Suggest Feature you want in AMD APP

Expose the BFI_INT instruction

The existing BFI_INT instruction is very powerful and speeds up some GPGPU tools by 10-30%. However developers have to use nasty hacks to make use of it because it is not exposed at all to CAL or OpenCL! This is a blatant mistake that could very easily be fixed by AMD. See how I hacked BFI_INT support in my whitepixel app: http://blog.zorinaq.com/?e=43

 

0 Likes
Marix
Adept II

Suggest Feature you want in AMD APP

Originally posted by: Meteorhead

Share GPUs across multiple threads

 

  Use all GPUs in the system concurrently from a single host thread



Aren't these features already supported by OpenCL?

0 Likes
himanshu_gautam
Grandmaster

Suggest Feature you want in AMD APP

mrbpix,

This is already being discussed among AMD developers. Thanks for reporting.

0 Likes
LeeHowes
Staff
Staff

Suggest Feature you want in AMD APP

 

Aren't these features already supported by OpenCL?

 

Yes, that gives a certain meaing to the word "easily". It's certianly to see areas where OpenCL has pushed CUDA development rather than the other way around.

0 Likes
bubu
Adept II

Suggest Feature you want in AMD APP

Add a watchdog enabled/disabled flag to the clGetDeviceInfo().

0 Likes
Meteorhead
Challenger

Suggest Feature you want in AMD APP

Originally posted by: LeeHowes


 

 

 

Aren't these features already supported by OpenCL?

 

 

 

 

Yes, that gives a certain meaing to the word "easily". It's certianly to see areas where OpenCL has pushed CUDA development rather than the other way around.

 

By "easily" I meant that it does not go against OCL specs (apart from C++ features). I do not know jack about making drivers threadsafe for using multiple gpu from single host thread, so I can only imagine how hard achieving such a feat is, just the same way I do not know how hard it is to redesign the linux driver.

I only know these things are not impossible, they are highly anticipated and NV has made them work. (Although many people anticipate proper OCL 1.1 support from driver side for NV cards)

0 Likes
LeeHowes
Staff
Staff

Suggest Feature you want in AMD APP

No, I mean it's really in the spec. 1.1 guarantees thread safety on almost all of the API calls. 1.0 did now, though the AMD implementation has always been implemented that way. Both multiple threads for one GPU and multiple GPUs for one thread have always worked in the AMD OpenCL implementation and are now specifieded to work official in OpenCL. Of course, that ignores limitations of the AMD implementation in working well with multiple GPUs that I'm aware people have complained about, but I'm not aware of any issues with thread safety of the API. 

I'm not sure why NVIDIA has been lagging a little about releasing official OpenCL 1.1 support, I'm sure it works well, though. I'm sure they will get there in the end as you say. 

0 Likes